Get the latest tech news
Even 'uncensored' models can't say what they want
A safety-filtered pretrain can duck a charged word without refusing. It puts a fraction of the probability an open-data pretrain puts there. We call that...
None
Or read this on Hacker News
